Does eIQ Neutron support Vision Transformer (RF-DETR / DETR) conversion for NPU acceleration on i.MX

vijayranaACL — Mon, 31 Aug 2026 12:22:26 GMT

Goal

Deploy RF-DETR Nano (vision transformer + DETR head) on NXP FRDM-IMX95
Run inference with eIQ Neutron NPU via libneutron_delegate.so
Need confirmation: Does eIQ support vision-transformer-based models for NPU acceleration?

ONNX export succeeds
TFLite conversion succeeds (~30 MB model)
CPU inference is correct — detections match PyTorch (dog/person/car on test image)
Model runs on FRDM with and without delegate

No NPU graph created after conversion
- Inspecting converted .tflite: no neutronGraph / NeutronOperator markers
- Neutron delegate appears to offload 0 nodes → full CPU fallback
Full INT8 quantization fails on transformer ops (RANGE, DIV, etc.)
Used dynamic-range quant instead (weights quantized, float32 I/O)

RF-DETR is transformer-heavy: attention, BatchMatMul, Softmax, LayerNorm, Gather, etc.
Neutron docs list mainly CNN ops (Conv2D, DepthwiseConv2D, Pooling, Add…)
- Neutron Converter
- Supported operators
Similar split-model approach needed on Rockchip NPU (backbone on NPU, decoder on CPU):
rfdetr-on-rockchip-npu

Does eIQ Neutron on i.MX95 support end-to-end vision transformer models (ViT / DETR / RF-DETR)?
If not, what is the recommended split (e.g. DINOv2 backbone on NPU, DETR head on CPU)?
Are BatchMatMul / Multi-Head Attention supported on Neutron-S for i.MX95 in current SDK?
Which eIQ Toolkit version matches BSP kernel 6.18.2-1.0.0 for FRDM-IMX95?
Should we use NXP tflite-profiler + tflite-quantizer instead of onnx2tf dynamic-range quant for better NPU compatibility?
- Model Deployment Quick Start
Any reference example for transformer or DETR-style models on i.MX95 Neutron

Host (convert):

neutron-converter \

--input rf_detr_nano_full_int8.tflite \

--output rf_detr_nano_neutron_imx95.tflite \

--target imx95

Board (inference):

python3 redetr_test_neutron_imx95.py \

--model rf_detr_nano_neutron_imx95.tflite \

--image dog.jpg \

--delegate /usr/lib/libneutron_delegate.so

Expected issue: converted model has no neutronGraph → no NPU acceleration.

Please confirm:

Is vision-transformer NPU acceleration supported on i.MX95 today?
If not, what is the official workaround (model split, different export path, or future roadmap)?